AI agent speedup is bounded by human judgment time, not model capability—teams maximize returns by encoding decisions as durable artifacts.

Gene Amdahl's 1967 law on parallel computing—that speedup is limited by the serial fraction—applies directly to multi-agent systems. The maximum speedup from AI agents is bounded by 1/H, where H is the fraction of workflow requiring human judgment (clarification, approval, taste calls, resolving ambiguity). If H = 40%, no agent improvement can exceed 2.5× speedup; at 50%, the ceiling is 2×.

The components that model improvements shrink—clarification and verification—are not the ones that dominate at scale. Once mechanical work is automated, taste and novel decisions remain largely irreducible by better models, making H itself the binding constraint rather than agent capability.

High-leverage investment is making human intervention 'self-liquidating': every human decision should produce an encoded artifact (test, spec update, documented decision) that prevents the same intervention recurrence. This requires 'configurancy'—explicit behavioral commitments, specs, conformance suites, and rationale—so agents operate autonomously where knowledge is already encoded.